Dual-input deep learning system for microbial identification from blood agar plates

Area of Science:

• Microbiology
• Computer Science
• Medical Diagnostics

Background:

• Microbial culture and colony morphology analysis requires specialized expertise, limiting automation in clinical microbiology.
• Accurate identification of pathogenic microbial species is crucial for effective patient treatment and infection control.

Purpose of the Study:

• To investigate the feasibility of automating pathogenic microbial species identification from microbial colony images.
• To develop and evaluate an automated system for microbial identification in clinical settings.

Main Methods:

• Two ResNet-50 based models were trained on distinct datasets: cropped colony images and tiled culture plate images.
• Models were trained using 10,048 colony and 23,003 tile images from 418 strains, with performance assessed via five-fold cross-validation.
• An ensemble model integrated outputs from both colony and tile image models for improved classification.

Main Results:

• The colony image model achieved 0.934 sensitivity and 0.993 specificity.
• The tile image model achieved 0.918 sensitivity and 0.991 specificity.
• The ensemble model demonstrated superior performance with 0.955 sensitivity and 0.995 specificity on independent strains.

Conclusions:

• The ensemble model approach offers high accuracy and robustness for microbial identification.
• This automated method has significant potential to aid clinical microbiology laboratories in pathogen detection.
• The study highlights the technical contribution of ensemble deep learning models to microbial diagnostics.